IrrNet: Spatio-Temporal Segmentation Guided Classification for Irrigation Mapping

Oishee Bintey Hoque; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 7983-7985

Irrigation mapping holds significant importance in modern agriculture and land management. Understanding and accurately mapping irrigation practices offer valuable insights into water resource utilization crop health and sustainable land use. By unraveling the interplay between spectral bands features and irrigation practices we can refine our models to better capture nuanced irrigation patterns even in regions with limited data availability. This not only improves our ability to monitor and manage water resources but also advances our understanding of the complex relationship between different image features. In this research proposal we present an innovative approach to comprehensive irrigation mapping using advanced deep learning techniques. Our hypothesis centers around the development of a robust model that leverages publicly available remote sensing data to map irrigation practices. Through the integration of attention mechanisms multi-task learning and conditional classification our proposed architecture aims to simultaneously detect field boundaries classify irrigation status and identify specific irrigation methods. We address the challenges of subtle feature variations and limited data availability by enhancing the model's focus on crucial aspects and incorporating synthetic data generation. By unifying these components we aspire to create an effective tool for understanding irrigation landscapes and expanding its utility to broader agricultural contexts.